International Journal of Heat and Mass Transfer 173 (2021) 121248

Contents lists available at ScienceDirect

International Journal of Heat and Mass Transfer

journal homepage: www.elsevier.com/locate/hmt

Review

Laminar natural convection in a square cavity with 3D random

roughness elements considering the compressibility of the fluid
Boqi Ren a, Chung-Gang Li a,b,∗, Makoto Tsubokura a,b
a
Graduate School of System Informatics, Kobe University, 1-1 Rokkodai, Nada-ku, Kobe, Hyogo 657 8501, Japan
b
Complex Phenomena Unified Simulation Research Team, RIKEN, Advanced Institute for Computational Science, 7-1-26 Minatojima-minami-machi, Chuo-ku,
Kobe, Hyogo 650-0047, Japan

a r t i c l e i n f o a b s t r a c t

Article history: Natural convection in a three-dimensional square cavity with random artificial roughness on both vertical
Received 26 October 2020 walls is studied numerically for a Rayleigh number of 106 . Based on consideration of realistic conditions,
Revised 9 March 2021
the roughness is generated using a given power spectrum density, and a compressible solver with precon-
Accepted 18 March 2021
ditioning that uses a dual time-stepping method is established to handle the low velocity of the natural
Available online 5 April 2021
convection flow. The compressible solver demonstrates satisfactory accuracy in terms of the average Nus-
Keywords: selt number when compared with incompressible benchmark solutions. However, the results for the local
Natural convection Nusselt number from the compressible and incompressible solvers differ. As the temperature difference
3D random roughness between the two vertical sidewalls increases, the maximum value of the local Nusselt number increases,
Compressible flow but in the downstream region, the local Nusselt number actually decreases. The cavity with rough side-
Nusselt number walls has a similar Nusselt number to the average Nusselt number of a hot sidewall in the case of a
smooth cavity. Because of the effects of the roughness on the thermal and velocity boundary layers, the
rough cavity case shows a better heat transfer performance than the smooth cavity case in the upstream
region, but in the downstream region, the heat transfer in the rough case becomes worse than that in
the smooth case.
© 2021 Elsevier Ltd. All rights reserved.

1. Introduction to use a compressible flow solver because of the conspicuous den-

In modern industry, many facilities operate under conditions
that involve high temperature differences, e.g., nuclear reactor
cooling, and the cooling systems used in construction and solar
collectors [1–3], so, in that case, the influence of compressibility
of fluids on the heat transfer performance becomes conspicuous.
On the other hand, to enhance the heat transfer, the roughness or
the presence of obstacles on the surfaces from which heat trans-
fer takes place has been researched widely, but most studies only
consider incompressible fluids [4–6]. Therefore, a study of the heat
transfer considering both roughness and the compressibility of flu-
ids is necessary and significant.
For the case of a high temperature difference (i.e., >30 K), Gray
and Giorgini [7] recommended considering the compressibility of
the fluid during numerical calculations, thus making it necessary
∗
Corresponding author at: Computational Fluid Dynamics Laboratory, Depart-
ment of Computational Science, Graduate School of System Informatics, Kobe Uni-
versity, 1-1 Rokkodai, Nada-ku, Kobe 657-8501, Japan.
E-mail addresses: bochiren1992@stu.kobe-u.ac.jp (B. Ren),
cgli@aquamarine.kobe-u.ac.jp (C.-G. Li), tsubo@tiger.kobe-u.ac.jp (M. Tsubokura).
sity variations of fluids. However, when compared with the speed
of sound, the magnitude of the fluid speed is too small to allow a
compressible solver to be used directly. For this reason, Weiss and
Smith [8] adopted a preconditioning method for application to nat-
ural convection in a horizontal annulus and their results showed
that the accuracy of this method was satisfactory under the condi-
tion where the Rayleigh number was equal to 4.7 × 104 . Yamamoto
et al. [9] simulated natural convection in a horizontal circular pipe
using the same method to deal with compressible flows at low ve-
locities and proved that the preconditioning method could achieve
good agreement with the experimental results in terms of the ther-
mal boundary layers and the Nusselt number under the conditions
of a temperature difference of 32.5 K and the Rayleigh number
Ra = 105 . In addition, the preconditioning method was shown to
accelerate the simulation toward convergence under steady-state
conditions [10].
Buoyancy driven flows in square cavities have previously been
studied widely. Bajorek and Lloyd [11] used experimental methods
to investigate the effects of rectangular partitions on the horizon-
tal walls in a square cavity on the heat transfer between the two
vertical walls. Their results showed that, when compared with a

https://doi.org/10.1016/j.ijheatmasstransfer.2021.121248
0017-9310/© 2021 Elsevier Ltd. All rights reserved.
B. Ren, C.-G. Li and M. Tsubokura International Journal of Heat and Mass Transfer 173 (2021) 121248

ber was 106 . However, there was an increment in the surface area
Nomenclature of 32% when using the single roughness element in comparison
with the smooth wall case. The mean Nusselt number of the case
L1 , L2 , L3 width, height and depth of the square cavity where two roughness elements were present showed a slight in-
A area of sidewalls crease in the mean Nusselt number (3.88% higher) when compared
C power spectral density (PSD) with the single roughness element case.
d distance from the cell center to the closest geom- In addition to the rectangular roughness elements, sinusoidal
etry element roughness elements have also been studied recently by many re-
g gravitational acceleration searchers. Bhardwaj et al. [13] simulated the flow field in a porous
h height of individual roughness element right-angled triangular cavity with a sine-wave-shaped left wall
k theermal conductivity and a smooth bottom wall to determine the effects of the fluctu-
Nu Nusselt number ating hot wall on the heat transfer and entropy generation prop-
P pressure erties of the cavity. Their results showed that, under the laminar
Pr Prandtl number flow condition, the undulation of the hot wall caused the strength
q wavevector of the vortex to be enhanced; in addition, both the average Nus-
R gas constant selt number and the maximum Nusselt number were higher than
Ra Rayleigh number based on cavity in the zero-undulation case. The increment in the maximum Nus-
height, gβ0 T L2 3 P r/ν0 2 selt number was 53% for a Rayleigh number of 106 and a Darcy
Rq root mean square roughness number of 0.01. However, in contrast, Yousaf and Usman [14] ob-
T temperature tained a lower average Nusselt number along a hot wall with si-
t physical time nusoidal roughness elements. At an Ra of 106 and using 10 rough-
u1 , u2 , u3 velocity components in x1 , x2 and x3 directions ness elements, the average Nusselt number decreased by 12.25%
x1 , x2 , x3 spatial coordinates when compared with that in the smooth wall case. Shahriari et al.
[15] also observed a reduction in the Nusselt number. At Ra = 105 ,
Greek symbols
with eight sinusoidal roughness elements on both hot and cold
β thermal expansion coefficient
walls in a square cavity, the Nusselt number decreased by approx-
T temperature difference, (TH − TC )
imately 20% when compared with that of the smooth wall case.
x 1 , x 2 , x 3 mesh size in x1 , x2 and x3 directions
In most previous experimental or numerical researches, sim-
t physical time step
ple geometric shapes, such as thin sheets and rectangular or si-
τ artificial time step
nusoidal shapes, were selected as the shapes of roughness ele-
δi j Kronecker delta
ments [12,14,15]. In contrast, to represent the roughness existing in
δx i central difference operator
nature, a three-dimensional artificial random roughness has been
γ adiabatic index
generated in the present study. The roughness elements are ran-
 preconditioning matrix
dom in shapes and heterogeneous in sizes. Furthermore, by tak-
θ nondimensional temperature,
ing the working conditions of practical industrial applications into
(T − Tc )/(Th − Tc )
account, a compressible solver with an all-speed Roe scheme has
ρ density
been established as the appropriate numerical method to investi-
τ artificial time
gate buoyancy-driven flows.
μ dynamic viscosity
ν kinematic viscosity 2. Physical model and governing equations
φ arbitrary physical value

Subscript Because the three-dimensional geometries of random rough

C cold wall surfaces cannot be ignored in this work, the numerical simulations
H hot wall of the flows in the square cavity are performed under 3D condi-
IC interface cell tions. Fig. 1 shows the physical model of current study with dif-
IP image point ferentially heated vertical sidewalls and adiabatic top and bottom
l local value surfaces. A high temperature of TH and a low temperature of TC ,
overall average value over an entire sidewall which are both constant and isothermal, are set on the left and
p primitive form of variables right sidewalls, respectively. The no-slip boundary condition is im-
w wall posed on the top, bottom and both sidewalls of the cavities, but
0 ambient or initial value slip boundary condition is imposed on both x3 -direction bound-
aries (at x3 /L3 = 0 and x3 /L3 = 1 planes). For all cavity flow, air
Superscript with a Prandtl number (Pr) of 0.72 is selected as the working fluid.
∗ nondimensional value The initial temperature of the air contained in the core of the
k artificial time square cavity is T0 = (TH + TC )/2 and the pressure is 101,300 Pa.
n physical time The artificial random surface roughness is generated based on
a given power spectrum density [16,17], as illustrated in Fig. 2. In
this research, the root mean square roughness
 (Rq) is equal to 1.5%
1  2
n
smooth cavity, the presence of the partitions caused the heat trans-
of the cavity width (L1 ), where Rq = n hi , and hi is the height
fer to be reduced by 12% and 21% at high Grashof numbers and low i=1
Grashof numbers, respectively. Shakerin et al. [12] used experimen- of each roughness element from the mean plane of the roughness.
tal methods to study the heat transfer from hot walls with differ- On the mean plane’s position, the heights of peaks and depths of
ent numbers of rectangular roughness elements within a square valleys all add to zero. Because of the random distribution of the
enclosure. For a single roughness element, the mean Nusselt num- roughness elements, the depth-to-width aspect ratio (L3 /L1 ) is se-
ber along the hot wall showed an increase of 12% when compared lected to be 2 and the arrangement of the roughness elements is
with the smooth wall under the condition that the Rayleigh num- symmetrical with respect to the middle of the depth of the cav-

2
B. Ren, C.-G. Li and M. Tsubokura International Journal of Heat and Mass Transfer 173 (2021) 121248

⎡ ⎤
ρUi
⎢ ρ ui u1 + pδi1 − μAi1 ⎥
⎢ ρ u u + pδ − μA ⎥
Fi = ⎢ i 2 i2 i2 ⎥, ∀i, j = 1, 2, 3, (3)
⎣ ρ ui u3 + pδi3 − μAi3 ⎦
(ρ e + p)ui − μAi j u j − k ∂∂xTi
and
⎡ 0
⎤
⎢ 0 ⎥
S = ⎢ −(ρ − ρ0 )g ⎥, (4)
⎣ ⎦
0
−(ρ − ρ0 )gu2
where Ai j = ∂ u j /∂ xi + ∂ ui /∂ x j − 2/3(∇ · u )δi j and δi j is the Kro-
necker delta. The relationship between the pressure and the den-
sity of the air is represented by the ideal gas equation, where
P = ρ RT . According to Sutherland’s Law [18], the viscosity (μ) and
the thermal conductivity (k) of the working fluid are determined
as follows:

T  32 T + 110.4
μ ( T ) = μ0 0
, (5)
T0 T + 110.4
Fig. 1. Simulation settings of the square cavity.
μ(T )γ R
k (T ) = , (6)
(γ − 1 )Pr
ity. As noted above, in many previous works, the existence of the
where ρ0 = 1.18 kg/m3 , g = 9.81 m/s2 , μ0 = 1.85 × 10−5 N · s/m2 ,
roughness will reduce the effective volume of fluid in the cavity.
T0 = 298.06 K, γ = 1.4, R = 287 J/kg, and P r = 0.72.
Therefore, to maintain a constant value for the effective volume,
rough surfaces have been installed on both hot and cold sidewalls
3. Numerical method
in this research, and the rough surface on the cold wall is gener-
ated by direct translation of the rough surface on the hot wall.
3.1. Compressible solver
To deal with the natural convection that occurs because of the
buoyancy force, the following compressible governing equation has
In natural convection, the magnitude of the fluid velocity due to
been considered:
the buoyancy force is generally smaller than the speed of sound by
∂ U ∂ F1 ∂ F2 ∂ F3 several orders of magnitude. For this reason, the original compress-
+ + + = S. (1) ible solvers are unsuited to analysis of this type of flow field. To
∂ t ∂ x1 ∂ x2 ∂ x3
increase the efficiency for solution of natural convection problems
U and Fi can be written in following format: involving compressible flows, a preconditioning method is used to
resolve the governing equations, which is developed by Weiss and
⎡ ρ ⎤ Smith [8]. When combined with the Roe scheme [19] and the dual
⎢ρ u 1 ⎥ time-stepping method [20], the new governing equation is shown
U = ⎢ρ u 2 ⎥, (2) as follows:
⎣ ⎦
ρ u3 ∂ Up ∂ U ∂ F1 ∂ F2 ∂ F3
ρe  + + + + = S, (7)
∂τ ∂ t ∂ x1 ∂ x2 ∂ x3

Fig. 2. 3D random artificial roughness and the power spectrum.

3
B. Ren, C.-G. Li and M. Tsubokura International Journal of Heat and Mass Transfer 173 (2021) 121248

where  is the preconditioning matrix provided by Weiss and

Smith, U p is the primitive form of the variables [P, u1 , u2 , u3 , T ],
and U is the conservative form of [ρ , ρ u1 , ρ u2 , ρ u3 , ρ e]. τ and t
represent the artificial and physical times, respectively.
The discretized form of Eq. (7) is

U pk+1 − U pk3U k+1 − 4U n + U n−1 1


 + + F k+1 − F k+1
τ 2t x1 1i+1/2, j,k 1i−1/2, j,k
1

 1


+ F2k+1 − F2k+1 + F3k+1 − F3k+1 = Sk . (8)
x2 i, j+1/2,k i, j−1/2,k x3 i, j,k+1/2 i, j,k−1/2

The iterations of the artificial time and the physical time are
indicated by the superscripts k and n shown in Eq. (8). It should
be noted that when the iteration of the artificial time reaches con-
vergence, the residual of ∂ U p /∂τ is less than 10−3 , and the mag-
nitude of the artificial term at the (k+1)th step is equivalent to
the magnitude of the physical time term at the (n+1)th step, then
Eq. (8) would then return approximately to the original Navier-
Stokes equation.
Eq. (8) can be expressed in the following format as:
 I 3 
+  −1 M +  −1 δx1 Akp + δx2 Bkp + δx3 C pk U p =  −1 Rk ,
τ 2t
(9)
where M = ∂ U/∂ U p , A p = ∂ F1k /∂ U p , = − Rk Sk
Fig. 3. Configuration of immersed boundary method.
(3U − 4U + U )/2t − (δx1 F1 + δx2 F2k + δx3 F3k ),
k n n −1 k and δxi is
the central difference operator. Xu et al. [21] recommended setting
the pseudo-time step, τ , to infinity then the Newton lineariza- 3.2. Immersed boundary method for compressible flows
tion error will be zero in each physical time step. Eventually,
Eq. (9) can be rewritten into Eq. (10): Because of the complex geometry of the random rough surface,
 3  the immersed boundary method(IBM) has been applied [23] to en-
 −1 M +  −1 δx1 Akp + δx2 Bkp + δx3 C pk U p =  −1 Rk , (10) sure that the geometrical details are reproduced as accurately as
2t
possible. The key point of IBM is to simulate the existence of ge-
The lower-upper symmetric-Gauss-Seidel (LUSGS) implicit ometry by imposing the physical values into the numerical grids in
method is then used to solve Eq. (10). the vicinity of geometry. Fig. 3 shows the configuration of IBM. The
It is helpful to divide the flux term given in Eq. (3) into an in- cells which are defined as interface cell (IC) has the characteristic
viscid term and a viscid term, as shown in Eqs. (11) and (12), re- that the distance (d) from the cell center to the closest geometry
spectively, to perform the calculation of Rk in Eq. (10). The inviscid element is less than one cell size. Image point (IP) is the point that
term in Eq. (11) can then be discretized into Eq. (13) using the Roe the distance from which to IC is equal to d in the normal direction
scheme. of the geometry element. IC and IP must be on the same side of
⎛ ρ ui ⎞ geometry. Then, according to the relative locations of surrounding
⎜ρ ui u1 + pδi1 ⎟ cells to IP, the physical value at IP, φIP , can be calculated by bilinear
Finviscid = ⎜ρ ui u2 + pδi2 ⎟, (11) interpolation:
⎝ ⎠
ρ ui u3 + pδi3 φIP = w1 φ1 + w2 φ2 + w3 φ3 + wIC φIC , (16)
(ρ e + p)ui
where w1 , w2 , w3 and wIP can be obtained using VanderMonde
⎛ ⎞ matrix [23]. Tullio et al.[24] recommended applying linear interpo-
0
⎜ μA i 1 ⎟ lation to evaluate φIC , shown in the following equations:
⎜ μA i 2 ⎟ 
Fviscous = −⎜ ⎟, (12) ∂φ φIP − φw
⎝ μA i 3 ⎠ φw = φIC − d = φIC − d, (17)
∂n 2d
μAij u j + k ∂∂xT i

1
Finviscid,i+1/2
1
= [FR (U ) + FL (U )] + Fd , (13)
φIC = (φw + φIP ), (18)
2
2
In present study, the no-slip and isothermal conditions are im-
where Fd is the Roe dissipation term.
plemented on the walls, so φw is an assigned value φassign , thus
For the two convective terms (FR and FL ) shown in Eq. (13), the
the Dirichlet condition should be considered. Then according to
values at the cell interfaces are calculated by interpolation of the
Eqs. (16)–(18), φIC can be presented as follows:
corresponding values from a group of cell centers through a fifth-
order MUSCL approach [33] without limiter functions, which are w1 φ1 + w2 φ2 + w3 φ3 + φassign
shown in the following way: φIC = Dirichlet condition, (19)
2 − wIC
1 However, for pressure, Neumann condition, which considers
UL1+1/2 = [−3Ui+2 + 27Ui+1 + 47Ui − 13Ui−1 + 2Ui−2], (14)
60 φw = φIP , should be used, then φIC can be presented as follows:

1 w 1 φ1 + w 2 φ2 + w 3 φ3
UR1+1/2 = [2Ui+3 − 13Ui+2 + 47Ui+1 + 27Ui − 3Ui−1 ]. (15) φIC = Neumann condition, (20)
60 1 − wIC + ε

4
B. Ren, C.-G. Li and M. Tsubokura International Journal of Heat and Mass Transfer 173 (2021) 121248

Fig. 4. Numerical grids distributions of smooth (a) and rough (b) cases at x3 = 0.5 plane.

where ε is a small value to prevent the denominator becoming Table 1

Convergence behaviors of hydrodynamic quantities and Nusselt numbers.
zero. Because when IC is very close to the wall, wIC will approach
1. Ra Ref. Numerical method u2 ∗ max (x1 /L1 ) Nu
In authors’ previous study [25], the natural convection of a 10 6
E Tric [27] Incompressible 0.258 (0.033) 8.877
heated sphere has been simulated through the IBM mentioned A Xu [28] Incompressible 0.258 (0.037) 8.881
above. The comparison of the averaged Nu shows good agreement Present Compressible (10 K) 0.260 (0.039) 8.854
with the result of Jia and Gogos [26], and the error is only 0.343%. Present Compressible (50 K) 0.259 (0.041) 8.874
Present Compressible (120 K) 0.264 (0.043) 8.787

3.3. Building cube method

numbers can be obtained using Eqs. (21) and (22). The average
The building cube method (BCM) is used to provide higher
Nusselt number over entire hot sidewall is calculated through
computation efficiency and better load balancing on a massively
Eq. (23). The maximum nondimensional vertical velocity at the
parallel computing system [25]. The BCM divides the calculation
center of the hot sidewall across the nondimensional width of the
domain into several cubes, and each cube contents 163 meshes. In
cavity (x1 /L1 ) is also presented
 in the table. The nondimensional
any individual cube, the mesh spacings along the three orthogonal
velocity is defined as u∗i = ui / gβ0 T L2 .
axes (directions x1 , x2 , and x3 ) are same, which allows the details
of the flow fields between two roughness elements to be captured 1 L2
Nu = ∫ N ul d x2 , (21)
more easily. L2 0

L2 ∂T
4. Results and discussion N ul = k (T ) , (22)
k0 (Th − Tc ) ∂ x1
4.1. Investigation of fluid compressibility in enclosed square cavity  A
1
Nuoverall = N ul dA, (23)
To discuss the effects of the compressible fluid on heat trans- A 0

fer, 3D cavities, shown in Fig. 1, with smooth hot and cold
sidewalls are investigated in this section. For all cavities, the
Rayleigh number remains constant, where the Rayleigh number
Ra = gβ0 T L2 3 P r/ν0 2 = 106 and all fluid properties are determined
based on the ambient temperature.
With regard to the influence of the compressibility of the work-
ing fluid on the heat transfer, three different temperature differ-
ences (10 K, 50 K, and 120 K) have been studied and the mesh
distributions have been shown in Fig. 4(a). The finest grids are im-
plemented in the vicinity of two sidewalls, and to save the compu-
tational resources, the coarser grids are implemented in the core
regions of the cavities. There are 256 finest grids along with the
height (L2 ) direction of the enclosed cavities, and the distributions
of grids along the depth (L3 ) direction are uniform for all three
cases. The Courant number (CFL) for all cases is set as 0.5 where
CF L = max(ui · t/xi ), i = 1, 2, 3 to represent the components
acting along the three axes in the coordinate system.
Table 1 shows the average Nusselt numbers across the x2 -
direction and the maximum values of the local Nusselt numbers
on the hot sidewalls across the height at x3 /L3 = 0.5 plane; these
where k0 = k(T0 ) and A = L2 · L3 .
The average Nusselt numbers of the three cases with various
temperature differences from Table 1 show good agreement with
the results from previous studies at the Ra of 106 . The maximum
deviation is 1.06%, which is the error between the case with the
temperature difference of 120 K and the results of A Xu [28].
Fig. 5 presents the local Nusselt number distributions at the mid-
dle depth on the hot sidewall. For the compressible solver, it is
evident that as T increases, the pattern of the local Nu numbers
on the hot wall becomes steeper, i.e., the maximum value of the
local Nu increases, but the cases with larger T values will have
smaller local Nu magnitudes in the x2 /L2 > 0.3 region. Between the
case with the temperature difference of 120 K and the results of
Kuyper et al. [29], the maximum local Nusselt numbers reach their
largest deviation of 6.233%. For the vertical velocity, Table 1 indi-
cates that as the temperature difference increases, the location of
the maximum vertical velocity moves away from the hot sidewalls.
This phenomenon where the distribution of the local Nusselt num-
ber changes with the temperature difference was also reported in
[30] and it was attributed to the temperature dependences of the

5
B. Ren, C.-G. Li and M. Tsubokura
Fig. 5. Local Nusselt numbers of smooth cases with various temperature differences.
Table 2
Values of average Nu for grid sensitivity study.
Mesh resolution Nuoverall Difference (%)
512 8.719 -
256 8.717 0.023
128 7.765 10.942
fluid properties, as shown in Eqs. (5) and (6). In summary, a tem-
perature difference between two heated sidewalls will change the
local Nusselt number distribution but will have no effect on the
average Nusselt number for a Rayleigh number of 106 . Therefore,
it becomes necessary to consider adopting a compressible solver
when local region heat transfer is the research target.
4.2. Grid sensitivity for the enclosure with rough sidewalls
Due to the complex 3D random roughness, the Immersed-
boundary method which is mentioned in Section 3.2 requires suit-
able girds to capture the geometric features of the roughness.
Meanwhile, for the simulation of the fluid among the roughness
elements, an appropriate mesh resolution is necessary. Therefore,
in this part, the cavity with rough sidewalls is filled by numeri-
cal grids with 3 different resolutions shown in Table 2. The mesh
resolutions of 128 and 512 are respectively generated by coarsen-
ing or refining the grids based on the resolution of 256, shown
in Fig. 4(b). For convenience, we use grid 128, grid 256 and grid
512 to represent these three mesh resolutions. Table 2 shows the
average Nusselt numbers over the entire hot sidewall from these
3 different grid strategies. After a comparison between the results
of grid 256 and grid 512, the average Nusselt numbers are almost
same and the difference is only 0.023%. However, the difference
between grid 512 and grid 128 is 10.942%.
Further, Fig. 6 shows the nondimensional temperature distri-
butions across the horizontal centerline of the cavities’ midplane
(x2 /L2 = 0.5 line at x3 /L3 = 0.5 plane). The profiles show that the
temperature distributions of grid 256 and gird 512 almost over-
lapped near the hot sidewall, but due to the insufficient resolution,
6
International Journal of Heat and Mass Transfer 173 (2021) 121248
Table 3
Maximum values of the local Nusselt numbers and
the average Nusselt numbers for the smooth and
rough cases.
Smooth Rough Difference (%)
Nulmax 18.134 60.207 +232.012
Nuoverall 8.874 8.717 -1.769
grid 128 shows a different profile of temperature from the other
two mesh resolutions. According to the arrangement of roughness
elements in current study, for the resolution of grid 256, there are
at least 4 grids for one peak or one valley on the rough surface,
but for the resolution of grid 128, this number is only 2 for one
peak or one valley.
Thus, all subsequent results in the following section are based
on the mesh resolution of grid 256.
4.3. Influence of roughness on heat transfer
Fig. 7 shows the local Nusselt number distributions on the
hot sidewalls of the smooth and rough cavity cases. For both the
smooth case and the rough case, the local Nusselt numbers show
a trend of being higher in the upstream region and lower in the
downstream region. Because of the existence of the peaks and val-
leys on the sidewalls, the local Nusselt number distribution on the
rough surface shows good agreement with the distribution of the
roughness elements and the local Nusselt numbers on the peaks
are larger than those in the valleys. Table 3 shows the average Nus-
selt numbers over the entire hot sidewall and the maximum values
of the local Nusselt numbers for the two cases, where the symbols
of + and – represent that compared with the results of the smooth
case, the increase or decrease of the investigated values due to the
rough sidewalls respectively. In Table 3, the largest local Nusselt
number in the rough cavity case is much larger than that in the
smooth case; however, in contrast, the average Nusselt number of
the rough case is slightly smaller than that in the smooth case by
1.769%.
B. Ren, C.-G. Li and M. Tsubokura International Journal of Heat and Mass Transfer 173 (2021) 121248

Fig. 6. Nondimensional temperature profiles.

Fig. 7. Distributions of the local Nu values of the (a) smooth and (b) rough cases on the hot walls.

Table 4 When compared with the smooth case, the local average Nus-
Local average Nusselt numbers in the upstream and downstream regions.
selt number in the upstream region in the rough case is higher,
Smooth Rough Difference (%) but in the downstream region, the local average Nusselt number
Upstream (x2 /L2 = 0 to 0.1) 17.613 18.622 +5.729 in the rough case is smaller than that in the smooth case. To ex-
Downstream (x2 /L2 = 0.75 to 1) 2.822 2.743 -2.799 plain this phenomenon in greater depth, the temperature distribu-
tion on a slice at a quarter of the depth of the cavity is shown in
Fig. 8, where the nondimensional temperature is calculated using
Table 4 presents the local average Nusselt numbers of the two the following equation: θ = (T − Tc )/(Th − Tc ). In the upstream re-
cases above in the upstream and downstream regions. Here, the gion, the peaks will reduce the thermal boundary layer thickness,
local average Nusselt number is considered to be the integral of the while the valleys can make the thermal boundary layer thicker,
Nusselt number on a local part of the heated sidewall, as defined which means that the thermal boundary layer will undergo drastic
in Eq. (24): changes in thickness within this region. However, as the thermal
boundary layer develops, the thickness will inevitably increase, and
1 Al then the effects of the roughness elements on the thermal bound-
Nul = ∫ N ul dA, (24)
Al 0 ary layer will decrease. From Fig. 8(c) and (d), the isothermal sur-
faces are distorted by the roughness elements near the rough side-
where the subscript l indicates a specific local part of the heated
sidewall.

7
B. Ren, C.-G. Li and M. Tsubokura
Fig. 8. Temperature distributions of smooth and rough cases: (a, b) slices at x3 /L3 =0.25 plane (c, d) isothermal surface in half cavity from x3 /L3 =0 to 0.5.
wall, but in the core region of the cavities, the temperature strati-
fications are almost identical for smooth and rough cases.
Fig. 9(a) and (b) show the nondimensional velocity magni-
tudes of the two cases on three different planes (x1 /L1 = 0.005,
x2 /L2 = 0.5 and x3 /L3 = 0.25) respectively. From this figure, when
the flow approaches the peaks on the rough surface, there is a
slight acceleration in the magnitude of the velocity when com-
pared with the smooth case, but after the peak has been crossed,
the ensuing valley will cause the fluid velocity to decelerate
sharply. Unlike the influence of the rough surface on the thermal
boundary layer, the influence of the rough surface on the magni-
tude of the velocity in the downstream region is obvious. This can
therefore be considered to be one explanation as to why the local
Nusselt number distribution presents the tendency where, in the
case with the rough heated sidewalls, the local Nusselt number is
higher than that in the smooth case in the vicinity of the peaks
8
International Journal of Heat and Mass Transfer 173 (2021) 121248
in the upstream region, while in the downstream region, the local
average Nusselt number is smaller than that in the smooth case.
Fig. 10 shows the contours of the nondimensional velocity mag-
nitude of both smooth and rough cases near the hot sidewall at
four different x1 -x2 planes (x3 /L3 = 0.125, 0.25, 0.375 and 0.5),
where the black curves are the profiles of hot sidewalls at these
planes. From the figure, the velocity magnitudes of the flow field in
the vicinity of the rough wall are smaller than smooth case gener-
ally, and only part of the fluid will be accelerated due to the rough-
ness. In the region where the shape of roughness elements does
not change drastically, the roughness plays a role more in slowing
down the velocity of adjacent fluid rather than accelerating it.
Fig. 11(a) and (b) show the velocity vector fields in the half re-
gion located close to the hot sidewall at the same location used in
Fig. 10(a) and (b). The figure containing the velocity vectors shows
no evident fluid circulation to be found in the area between the
two roughness elements on the x1 -x2 plane at Ra = 106 . In addi-
B. Ren, C.-G. Li and M. Tsubokura International Journal of Heat and Mass Transfer 173 (2021) 121248

Fig. 9. Nondimensional velocity magnitudes (where (a) and (b) show the smooth and rough cases, respectively) and vectors (where (c) and (d) show the smooth and rough
cases, respectively).

Fig. 10. Contours of Nondimensional velocity magnitude of smooth (a) and rough (b) cases at different x1 -x2 planes.

9
B. Ren, C.-G. Li and M. Tsubokura International Journal of Heat and Mass Transfer 173 (2021) 121248

Fig. 11. Velocity vector fields of smooth (a) and rough (b) cases at different x1 -x2 planes.

Fig. 12. Nondimensional temperature gradients across the height of the cavity at two different measurement locations (black curves: profiles of the rough surface at the two
measurement locations).

tion, the eddies that occur near the core regions of the two cavities
do not show particularly notable differences.
In Fig. 12, the components of the nondimensional temperature
gradients in the x1 -direction have been measured in the cavities
with both rough and smooth sidewalls at two measurement loca-
tions. These two locations are two different lines parallel to the x2 -
direction(x3 /L3 = 0.55 line and x3 /L3 = 0.8 line) at x1 /L1 = 0.05
plane starting from the mean plane of the hot rough sidewall. The
black curves shown in Fig. 12 are the profiles of the hot rough
surfaces at x1 -x2 planes where two measurement locations lie re-

10
B. Ren, C.-G. Li and M. Tsubokura
Fig. 13. Nondimensional temperatures at the center of the adiabatic surface across the cavity height.
Fig. 14. Magnitudes of the velocity component in the x3 direction.
spectively. The figure shows that in the upstream regions, the tem-
perature gradient fluctuates dramatically and this trend follows
the distribution of the roughness elements perfectly, which means
that higher peaks will bring increased heat transfer. However, this
dominance of the roughness decreases in the end region, partic-
ularly in the downstream region, and after the nondimensional
height exceeds 0.8, the patterns from the two cases almost over-
lap. In Fig. 13, at the center of the adiabatic surface and across
the heights of the cavities, the temperatures in both the rough
and smooth cases have been measured. The temperature curves
show very good agreement, thus indicating that the random ar-
tificial roughness cannot influence the heat transfer that occurs in
the core region of the square cavity.
Figs. 14 and 15 show the effects of the surface roughness on
the velocity component in the x3 -direction. The slices in Fig. 14 are
located at different x1 -x2 planes. The magnitudes of this veloc-
ity component shows a trend where the initial magnitude is very
small in the upstream region and then increases to reach a max-
imum in the middle region; however, in the downstream region,
deceleration occurs. Furthermore, the effect of the rough surface
cannot reach the core region of the cavity. The slice shown in
Fig. 15 is located at a position of 0.5% of the cavity width from the
11
International Journal of Heat and Mass Transfer 173 (2021) 121248
mean plane of the rough surface. In this figure, the changes in the
colors near the roughness elements and the speed vectors demon-
strate that the geometry of the 3D random roughness structure
gives the fluid the possibility of bypassing the roughness elements
and no evident eddies were found. However, in two-dimensional
cavity flow with homogeneous roughness elements on the cavity
sidewalls [15,16], the fluid will become trapped more easily be-
tween two adjacent roughness elements, and this represents a ma-
jor cause of the reduction in the heat transfer.
5. Conclusions
Natural convection in a cavity with 3D random roughness ele-
ments on its sidewalls has been investigated at a Rayleigh number
of 106 using a compressible solver. The roughness was generated
using a given power spectrum density and the temperature differ-
ence between the two sidewalls was 50 K. The distribution of the
local Nusselt number on the hot sidewall showed that cavities with
both rough and smooth sidewalls followed a trend where the local
Nusselt number was higher in the upstream region and lower in
the downstream region. The maximum value of the local Nusselt
number for the rough surface is 2.32 times larger than that in the
B. Ren, C.-G. Li and M. Tsubokura
Fig. 15. Flow field among the roughness elements of the hot sidewall.
smooth case, but the average Nusselt number is 1.8% smaller than
that in the smooth case. The peaks on the rough surface will re-
duce the thermal boundary layer thickness but the corresponding
valleys will increase this thickness in the upstream region. How-
ever, in the downstream region, the thermal boundary layer thick-
ness is thick enough to ensure that this effect is no longer obvious.
Near the hot sidewall, the peaks accelerate the flow field slightly,
but the valleys cause a conspicuous deceleration of the flow field;
this phenomenon exists in both the upstream and downstream re-
gions. Therefore, the Nusselt number on the rough surface in the
upstream region shows a larger average value than that of the
smooth case in the same region, but will show a smaller average
value in the downstream region. However, the fluid will bypass the
elements in the 3D simulations, rather than become trapped be-
tween two roughness elements, and this is one reason why the av-
erage Nusselt number in the rough case is not much smaller than
that in the smooth case.
Declaration of Competing Interest
The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared to
influence the work reported in this paper.
Acknowledgement
This work was supported by Japan Society for the Promotion of
Science (JSPS) KAKENHI [Grant number JP19K14890].
References
[1] I. Amber, T.S. O’Donovan, Natural convection induced by the absorption of so-
lar radiation: a review, Renew. Sustain. Energy Rev. 82 (2018) 3526–3545.
[2] Y. Ma, Z. Yang, Simplified and highly stable thermal Lattice Boltzmann method
simulation of hybrid nanofluid thermal convection at high Rayleigh numbers,
Phys. Fluids 32 (1) (2020) 012009.
[3] M.R. Amin, Natural convection heat transfer in enclosures fitted with a peri-
odic array of hot roughness elements at the bottom, Int. J. Heat Mass Transf.
36 (3) (1993) 755–763.
12
International Journal of Heat and Mass Transfer 173 (2021) 121248
[4] X. Zhu, R.J.A.M. Stevens, R. Verzicco, et al., Roughness-facilitated local 1/2 scal-
ing does not imply the onset of the ultimate regime of thermal convection,
Phys. Rev. Lett. 119 (15) (2017) 154501.
[5] S. Wagner, O. Shishkina, Heat flux enhancement by regular surface roughness
in turbulent thermal convection, J. Fluid Mech. 763 (2015) 109.
[6] B.N. Prasad, J.S. Saini, Effect of artificial roughness on heat transfer and friction
factor in a solar air heater, Sol. Energy 41 (6) (1988) 555–560.
[7] D.D. Gray, A. Giorgini, The validity of the Boussinesq approximation for liquids
and gases, Int. J. Heat Mass Transf. 19 (5) (1976) 545–551.
[8] J.M. Weiss, W.A. Smith, Preconditioning applied to variable and constant den-
sity flows, AIAA J. 33 (11) (1995) 2050–2057.
[9] S. Yamamoto, D. Niiyam, B.R. Shin, A numerical method for natural convection
and heat conduction around and in a horizontal circular pipe, Int. J. Heat Mass
Transf. 47 (26) (2004) 5781–5792.
[10] E. Turkel, Preconditioning techniques in computational fluid dynamics, Annu.
Rev. Fluid Mech. 31 (1) (1999) 385–416.
[11] S.M. Bajorek, J.R. Lloyd, Experimental investigation of natural convection in
partitioned enclosures, J. Heat Transf. 104 (3) (1982) 527–532.
[12] S. Shakerin, M. Bohn, R.I. Loehrke, Natural convection in an enclosure with
discrete roughness elements on a vertical heated wall, Int. J. Heat Mass Transf.
31 (7) (1988) 1423–1430.
[13] S. Bhardwaj, A. Dalal, S. Pati, Influence of wavy wall and non-uniform heat-
ing on natural convection heat transfer and entropy generation inside porous
complex enclosure, Energy 79 (2015) 467–481.
[14] M. Yousaf, S. Usman, Natural convection heat transfer in a square cavity with
sinusoidal roughness elements, Int. J. Heat Mass Transf. 90 (2015) 180–190.
[15] A Shahriari, E.J. Javaran, M. Rahnama, Effect of nanoparticles Brownian motion
and uniform sinusoidal roughness elements on natural convection in an enclo-
sure, J. Therm. Anal. Calorim. 131 (3) (2018) 2865–2884.
[16] M.M. Kanafi, A.J. Tuononen, Top topography surface roughness power spectrum
for pavement friction evaluation, Tribol. Int. 107 (2017) 240–249.
[17] Y. Xian, P. Zhang, S. Zhai, et al., Re-estimation of thermal contact resistance
considering near-field thermal radiation effect, Appl. Therm. Eng. 157 (2019)
113601.
[18] J.D. Anderson Jr., Hypersonic and High-Temperature Gas Dynamics, American
Institute of Aeronautics and Astronautics, 2006.
[19] P.L. Roe, Approximate Riemann solvers, parameter vectors, and difference
schemes, J. Comput. Phys. 43 (2) (1981) 357–372.
[20] S. Venkateswaran, C. Merkle, Dual time-stepping and preconditioning for un-
steady computations, in: 33rd Aerospace Sciences Meeting and Exhibition,
AIAA, Reno, NV, 1995, p. 78.
[21] X. Xu, J.S. Lee, R.H. Pletcher, A compressible finite volume formulation for large
eddy simulation of turbulent pipe flows at low Mach number in Cartesian co-
ordinates, J. Comput. Phys. 203 (1) (2005) 22–48.
[23] R. Ghias, R. Mittal, H. Dong, A sharp interface immersed boundary method for
compressible viscous flows, J. Comput. Phys. 225 (1) (2007) 528–553.
[24] M.D. de Tullio, P. De Palma, G. Iaccarino, et al., An immersed boundary method
for compressible flows using local grid refinement, J. Comput. Phys. 225 (2)
(2007) 2098–2117.
B. Ren, C.-G. Li and M. Tsubokura
[25] C.G. Li, M. Tsubokura, R. Bale, Framework for simulation of natural convection
in practical applications, Int. Commun. Heat Mass Transf. 75 (2016) 52–58.
[26] H. Jia, G. Gogos, Laminar natural convection heat transfer from isothermal
spheres, Int. J. Heat Mass Transf. 39 (8) (1996) 1603–1615.
[27] E. Tric, G. Labrosse, M. Betrouni, A first incursion into the 3D structure of nat-
ural convection of air in a differentially heated cubic cavity, from accurate nu-
merical solutions, Int. J. Heat Mass Transf. 43 (21) (20 0 0) 4043–4056.
[28] A. Xu, L. Shi, H.D. Xi, Lattice Boltzmann simulations of three-dimensional ther-
mal convective flows at high Rayleigh number, Int. J. Heat Mass Transf. 140
(2019) 359–370.
13
International Journal of Heat and Mass Transfer 173 (2021) 121248
[29] R.A. Kuyper, T.H. Van Der Meer, C.J. Hoogendoorn, et al., Numerical study of
laminar and turbulent natural convection in an inclined square cavity, Int. J.
Heat Mass Transf. 36 (11) (1993) 2899–2911.
[30] T. Fusegi, J.M. Hyun, Laminar and transitional natural convection in an enclo-
sure with complex and realistic conditions, Int. J. Heat Fluid Flow 15 (4) (1994)
258–268.
[33] Hong Kim Kim, Chongam Kim, Accurate, efficient and monotonic numerical
methods for multi-dimensional compressible flows: Part II: Multi-dimensional
limiting process, Journal of Computational Physics 208 (2) (2005) 570–615.